A detector cannot know which software was open while you typed. It reads the finished text and estimates how predictable the word choices are. Using autocomplete does not identify the text as machine output, but detectors can still misclassify human writing.

What a detector actually measures

These tools look at statistical properties of the text: how predictable each word is given the ones before it, how much the sentence lengths vary, how wide the vocabulary is. Text generated by a language model tends to sit in the high-probability middle of those distributions, because that is what the model is built to produce. Human writing wanders more.

That is a judgement about the text, not a record of how it was produced. Nothing in a document says which keystrokes came from a suggestion — and equally, nothing protects writing that happens to be plain and regular.

The research is not kind to detectors

An independent evaluation by Liang, Yuksekgonul, Mao, Wu and Zou at Stanford was published in Patterns in 2023. They ran seven widely used detectors over 91 TOEFL essays written by people and 88 essays by US eighth-graders, also written by people.

  • The eighth-grade essays were classified accurately — about 5.1% were wrongly flagged.
  • The TOEFL essays were not: the average false-positive rate was 61.3%.
  • All seven detectors agreed in wrongly flagging 19.8% of the TOEFL essays.
  • At least one detector flagged 97.8% of them.

The authors attribute this result to the features the detectors use. Limited vocabulary range and predictable sentence structure can appear machine-like to a detector, but they are also common in writing produced in a second language. The tools therefore risk penalising human writing for its statistical plainness.

That study is from 2023 and detectors have changed since. What has not changed is the shape of the problem: these are probabilistic classifiers, and none of them publishes a per-document error rate you could rely on to accuse someone.

So where does autocomplete sit

It depends entirely on how much of the text is yours.

What you did What the detector sees
Accepted a few word completions in sentences you composedYour writing, with your structure and vocabulary
Accepted a phrase now and then to finish a routine sentenceMostly your writing; routine sentences are routine either way
Asked a chatbot to write the paragraph and pasted itGenerated text, because it is generated text
Wrote it yourself, plainly, in a second languagePossibly flagged anyway — see above

The relevant distinction is whether you composed the text yourself. Finishing your own sentence with a suggestion retains your composition, while asking for and accepting a generated paragraph does not.

If you are being assessed on originality

  • Find out the actual rule. Many institutions permit spelling, grammar and predictive-text assistance and prohibit generated content. Some prohibit everything. A written rule is worth more than an assumption in either direction.
  • Keep your version history. Google Docs and Word both keep one. A document that grew over hours, with additions, deletions and sentences written twice, is a record of the writing happening — which is a different kind of evidence from a score, and one a score cannot answer.
  • Do not rewrite to beat a detector. Padding sentences with unusual words to lower a score makes your writing worse and is indistinguishable from what a “humanizer” tool does. If you are accused, the version history is the answer.
  • Treat a score as a starting point for a conversation, not a verdict. Given false-positive rates like those above, that is all it can honestly be.

Where a local autocomplete differs

One practical difference has nothing to do with detection. Cloud writing assistants send what you type to a server to generate their suggestions. Typeahead runs a compact model on your own PC through llama.cpp, so once a model is downloaded, nothing you write is transmitted anywhere and it works with no connection at all.

It suggests the next few words of the sentence you are already writing, next to the caret, and inserts nothing unless you press Tab. It does not generate paragraphs, and it has no “write this for me” function — which is the feature that would actually put generated text in your document.

None of that guarantees a particular detector result. Your sentences remain on your machine, and the tool does not include a function that generates paragraphs for you.

Sources

Frequently asked questions

Can an AI detector tell that I used autocomplete?
No. Detectors analyse the finished text for statistical patterns. They have no record of which software was running or which keystrokes came from a suggestion.
How accurate are AI detectors?
Unreliable enough that a score should not be treated as proof. In a Stanford evaluation published in Patterns in 2023, seven detectors wrongly flagged an average of 61.3% of human-written TOEFL essays, while wrongly flagging about 5.1% of essays by US eighth-graders.
Why do detectors flag non-native English writers so often?
They treat limited vocabulary range and predictable sentence structure as machine-like, and those are the characteristics of competent writing in a second language. The bias is in what the detectors measure.
Is using predictive text considered cheating?
That depends on the institution’s written rule, and the rules differ. Many draw the line at generated content rather than at typing assistance. Ask rather than assume, in either direction.
What should I do if I am wrongly accused?
Show the version history of the document. Google Docs and Word both record how a file was built up over time, and an edit trail is a record of the writing actually happening — which is something a probability score cannot speak to.
Does Typeahead write paragraphs for me?
No. It suggests the next few words of the sentence you are typing and inserts them only if you press Tab. There is no command that generates a paragraph.